discussion papers in loans, ordering and shortage costs in start … · 2012. 3. 26. · loans,...
TRANSCRIPT
Discussion Papers in
Management
Loans, ordering and shortage costs in start-ups: a dynamic stochastic decision approach Edgar Possani School of Management University of Southampton Lyn C Thomas School of Management University of Southampton and Thomas W Archibald School of Management University of Edinburgh
Number M02-8 June 2002 ISSN 1356-3548
Loans, ordering and shortage costs in start-ups:a dynamic stochastic decision approach
Edgar Possani
School of Management
University of Southampton
SO17 1BJ
Southampton, UK
Lyn C. Thomas
School of Management
University of Southampton
SO17 1BJ
Southampton, UK
Thomas W. Archibald
School of Management
University of Edinburgh
50 George Square
EH8 9JY
Edinburgh, UK
Loans, ordering and shortage costs in start-ups:
a dynamic stochastic decision approach
Abstract
Start-up companies are a vital ingredient in the success of a globalised networked
world economy. We believe such companies are interested in maximising the chance of
surviving in the long-term. We present a markov decision model to analyse survival
probabilities of start-up manufacturing companies. Our model examines the implica-
tions of their operating decisions. In particular, it examines their inventory strategy
including purchasing, shortage and ordering costs, as well as the in uence of loans on
the �rm. It is shown that although the start-up company should be more conservative
in its component purchasing strategy than if it were a well established company it
should not be too conservative. Nor is its strategy monotone in the amount of capital
available.
Keywords:
Inventory control, dynamic programming, risk analysis, manufacturing.
Loans, ordering and shortage costs in start-ups:
a dynamic stochastic decision approach
Introduction
Start-up companies are a vital ingredient in the success of a globalised networked
world economy. It is important to identify strategies that ensure success for these
newly formed companies. In this paper we analyse the relationship between �nance
and production costs in such an environment. Start-up companies di�er from those
that are well-established as shown by Archibald et al1. The model we consider is an
extension on the one they introduced. The main idea is that a newly created �rm
is interested primarily on the probability of survival as opposed to pro�ts being their
sole objective. Strategies that look to maximise the average pro�t are suited to well
established companies. We show how these strategies di�er for a start-up company, and
how as certain quantities of capital are reached it behaves more like an established �rm.
We focus on the inventory strategy of the �rm, and analyse its behaviour with changes
in the �rms capital, inventory on stock, ordering costs, purchasing costs, shortage costs
and loans.
There is little work done on making joint production and �nancial decisions for
(start-up) companies. Traditionally, �nance and production decisions have been taken
separately. This is largely due to the Modigliani-Miller2 proposition, that the deci-
sions could be taken independently under perfect and competitive markets. However,
in reality markets are imperfect (there are tax advantages of debt, bankruptcy cost,
etc.), and as Cifarelli et al3 mention, it has led to overlooking some of the dynamic
implications of the structure of capital. They study stochastic models to analyse the
value of a �rm over time, under di�erent �nancing policies, and are interested in the
probability of bankruptcy and repayment of loans. However, their analysis is not for
start-up companies, and they do not consider the in uence the ordering policy has on
the probability of survival of the �rm.
Buzacott and Zhang4 is one of the few papers that does analyse the relationship
between borrowing and production decisions in start-up companies. Their model con-
siders asset-based �nancing. For start-up companies the �nancing comes from loans
taken on the inventory of the company, as they lack the �x assets and �nancial re-
1
sources of an established company. They present a linear programming formulation to
express the ow of inventories, raw material and customer payments over time, to �nd
which production policy is convenient. They compare asset-based loans with unsecured
loans, and suggest that the largest impact on the �rm's pro�tability comes from the
bank's assessment of the risk associated with the �rm's WIP inventory. Finally they
analyse a periodic review myopic policy. They �nd the optimal ordering quantity for
di�erent inventory and equity (retained pro�ts) levels. They point out that the myopic
solution is not always adequate, and suggest the use of dynamic programming models
to aid the �rm's decision making. This supports the analysis that we do in this pa-
per, but they do not consider explicitly the inventory decision under the probability of
survival objective. Freel5 overviews the state of �nancing in product innovating small
manufacturing �rms within the UK.
In the following section, we present the survival probability model we analyse for a
start-up company and its properties. We show how the probability of survival behaves
in a similar way to how it behaved in the model presented by Archibal et al1, highlight-
ing the properties for the extensions. In the next section we give some examples of the
model's behaviour comparing di�erent levels of shortage costs, ordering costs, and bank
loans. We analyse the average reward model for a mature company, and prove that
the optimal inventory policies are more cautious for the survival model than for the
pro�t maximising model. Our conclusions are summarised in the last section, where
we outline some interesting areas of further work.
Survival Probability Model
In this section we present the Markov Decision model focusing on the survival proba-
bility of the start-up company. This model is an extension on the model presented by
Archibald et al1, which did not include either ordering or shortage costs nor loans on
assets. As pointed out by Buzacott and Zhang4 in the case of a manufacturing start-up
company the main asset are its inventories hence, loans taken on them are appropriate
for such companies. We do not assume that the demand for the product is known as
they did.
The simpli�ed model we study is as follows. The manufacturer makes one type of
unit from a single type of components that has to be purchased from other manufac-
2
turers. We consider the lead time for ordering as �xed, and taken as 1 time period.
The demand for the unit is random with independent identical distributions each time
period. If the demand in a period exceeds the number of components available in stock
then the excess demand is lost at a shortage cost of r. The shortage cost we consider
is not a stock out cost but rather the administrative cost involved in not supplying the
demand. The decision the �rm faces is how many components to order each period.
Figure 1 shows a time line for the events in one period. We assume that the bank
checks the �nances of the �rm at the end of each period. We pay the supplier on
receiving the components, and not on ordering. The components are delivered at the
beginning of the next period (constant lead time of one unit). We assume units cannot
be manufactured if components are not available. These assumptions really mean that
the manufacturing time and any variation in lead time is small compared with the lead
time itself. Ordering too many components ties up the �rm's capital in stock which is
not required, whereas ordering too few leads to unsatis�ed demand and hence a loss of
pro�t.
Figure 1: A time line for the events in one period
Let S be the selling price of each unit;
C be the cost of buying a single component;
H be �xed overhead costs per period (e.g. cost of sta� and premises) which are
incurred irrespective of the activity of the �rm;
w be the cost of submitting an order for components;
r be the shortage cost per unit of unsatis�ed demand;
3
p(d) the probability of there being a demand of d units;
M = maxfdjp(d) > 0g be the maximum possible demand that can be satis�ed
in a period (this can be interpreted as the maximum production capacity in a
period), and
d =P
M
d=0dp(d) be the average demand per period.
We de�ne q(n; i; x) as the maximum probability that a �rm will survive n more
periods given that it has i components in stock and x units of capital. We assume that
storage constraints put some upper limit on i and thus we have a �nite action space
since we cannot order more than this amount. q(n; i; x) is the optimal function for a
�nite horizon dynamic programming problem with a countable state space (x assumed
to have discrete levels), and a �nite action space - the amount k to order. Thus it has
an optimal non-stationary policy, see Puterman6. Moreover, the survival probability
q(n; i; x) satis�es the following dynamic programming optimality equation
q(n; i; x) = maxk
n MXd=0
p(d) q(n� 1; i+ k �min(i; d); x+ Smin(i; d) (1)
� Ck � rmax(0; d� i)�min(w; kw)�H)o:
As extra cash is available through the borrowing on the inventory, the bankruptcy
line for this model di�ers from the one presented by Archibald et al1. The loan on the
assets (inventory) is made explicit in the boundary conditions where q(n; i; x) = 0, if
x + �Ci < 0, and q(0; i; x) = 1 for i � 0, and x + �Ci � 0, where � is the proportion
of the colateral that the bank uses in setting the loan. The loan we are considering is
more of a variable overdraft, as no interest rate is charged for it in equation (1). Note
that 0 � � < 1. Recall that C was the purchasing price of one unit of component, thus
Ci is the value of the inventory on stock, and �Ci is the valuation by the bank of the
inventory i on stock, that is, the loan the �rm would receive at the beginning of the
period. If the �rm goes bankrupt (i.e. if at a given period x+ �Ci < 0) any inventory
in stock is taken by the bank.
Note that min(w; kw) is the cost of making an order. When the order quantity is
k > 0, min(w; kw) = w is the cost of ordering components, and when no components are
ordered, k = 0, min(w; kw) = 0, there is no ordering cost. For each unit of unsatis�ed
demand max(0; d� i) there is a cost of r, thus rmax(0; d� i) is the shortage cost. If the
4
demand is smaller than the inventory available then there is no shortage cost, however,
if the demand in a given period is higher than can be supplied from the inventory in
stock (d > i) then we incur a cost of r(d� i). If there are no shortage costs, r = 0, no
ordering costs, w = 0, and no loans � = 0, we are left with the original model studied
by Archibald et al1.
We de�ne k(n; i; x) = argmaxfP
M
d=0 p(d) q(n � 1; i + k � min(i; d); x + Smin(i; d)
� Ck � rmax(0; d� i) � min(w; kw)� H)g; the optimal action in state (n; i; x) for
given inventory level i and capital x.
Properties
There are some obvious properties which one expects of the survival probability q(n; i; x)
as n, i, x varies, that are similiar to the ones obtained in Archibald et al1. We also ex-
pect the probability of survival not to increase with higher shortage and ordering cost,
or with lower valuations of the stock from the bank. These properties are summarised
in Lemma 1, and Theorem 1.
Lemma 1
i) q(n; i; x) is non-increasing in n.
ii) q(n; i; x) is non-decreasing in x.
iii) q(n; i; x) is non-decreasing in i.
Proof:
The proof is by induction on n. Since q(n; i; x) = 0 when x + �Ci < 0 for all n,
q(0; i; x) = 1 when x + �Ci � 0 for i � 0 and q(1; i; x) � 1 when x + �Ci � 0 and
i � 0, all three hypotheses hold in the case n = 0.
Assume all three hypotheses hold for n, and use maxi
faig�maxi
fbig � maxi
fai� big
to show for given values of r; w � 0, and � � 0 that
i) q(n+ 1; i; x)� q(n; i; x) �
maxk
fMPd=0
p(d)( q(n; i + k �min(i; d); x + Smin(i; d)� kC
� rmax(0; d� i)�min(w; kw)�H)
� q(n� 1; i+ k �min(i; d); x+ Smin(i; d)� kC
�rmax(0; d� i)�min(w; kw)�H) )g � 0:
Hence, hypothesis (i) holds for n+ 1.
5
ii) q(n+ 1; i; x)� q(n+ 1; i; x+ a) �
maxk
fMPd=0
p(d)( q(n; i+ k �min(i; d); x+ Smin(i; d)� kC � rmax(0; d� i)
�min(w; kw)�H)
�q(n; i + k �min(i; d); x + a+ Smin(i; d)� kC � rmax(0; d� i)
�min(w; kw)�H) )g � 0 (where a > 0):
Hence, hypothesis (ii) holds for n+ 1.
iii) q(n+ 1; i; x)� q(n+ 1; i+ 1; x) �
maxk
fi�1Pd=0
p(d)( q(n; i+ k � d; x+ Sd� kC �min(w; kw)�H)
�q(n; i + 1 + k � d; x+ Sd� kC �min(w; kw)�H) )
+p(i)( q(n; k; x+ Sd� kC �min(w; kw)�H)
�q(n; 1 + k; x+ Sd� kC � r �min(w; kw)�H) )
+MP
d=i+1
p(d)(q(n; k; x+ Si� kC �min(w; kw)� r(d� i)�H)
�q(n; k; x+ Si + S � kC �min(w; kw)� r(d� i)�H))g
� maxk
fiP
d=0
p(d)( q(n; i+ k � d; x+ Sd� kC �min(w; kw)�H)
�q(n; i+ 1 + k � d; x+ Sd� kC �min(w; kw)�H) )
+MP
d=i+1
p(d)(q(n; k; x+ Si� kC �min(w; kw)� r(d� i)�H)
�q(n; k; x+ Si + S � kC �min(w; kw)� r(d� i)�H))g � 0:
Hence, hypothesis (iii) holds for n+ 1.
�
Theorem 1
i) q(n; i; x) is non-decreasing in �.
ii) q(n; i; x) is non-decreasing in S.
iii) q(n; i; x) is non-increasing in C.
iv) q(n; i; x) is non-increasing in H.
v) q(n; i; x) is non-increasing in r.
vi) q(n; i; x) is non-increasing in w.
Proof:
i) The proof is by induction on n. Since q(n; i; x) = 0 when x + �Ci < 0 for all n,
q(0; i; x) = 1 when x+�Ci � 0 for i � 0 and q(1; i; x) � 1 when x+�Ci � 0 and i � 0,
the hypothesis holds in the case n = 0. Assume the hypothesis holds for n, and that
�1 � �2 while values r; w � 0 are �xed. Let q0(n; i; x; �) be the survival probability
6
when � is the colateral fraction assumed by the bank. Use maxifaig � maxifbig �
maxifai � big to show that:
q0(n+ 1; i; x; �1)� q
0(n+ 1; i; x; �2) �
maxk
nPM
d=0 p(d) q0(n; i+ k �min(i; d); x+ Smin(i; d)
� rmax(0; d� i)� Ck �min(w; kw)�H;�1)
�q0(n; i+ k �min(i; d); x+ Smin(i; d)
� rmax(0; d� i)� Ck �min(w; kw)�H;�2)o� 0
Hence, hypothesis (i) holds for n+ 1.
ii) Note that in equation (1) there is more capital available with a sale price S1 than
with a sale price S2 when S1 > S2, hence by Lemma 1 ii), the property is valid.
iii)- vi) Note that in equation (1) there is less capital available under component cost
C1 than component cost C2 when C1 > C2, and hence by Lemma 1 ii), the property
(iii) is valid. We can use the same argument for the overhead costs, shortage costs and
ordering costs separately to justify properties iv), v), and vi), as they have the same
e�ect in capital.
�
Lemma 1, and Theorem 1 con�rm that a �rm is better o� with more inventory,
capital, and higher loans, or with smaller ordering, and shortage costs. On the other
hand the probability of surviving cannot increase as the time horizon increases.
It is of interest to know for what values of x there is no chance of survival and for
what x the survival is certain. If p(0) > 0 one extreme is a continual zero demand for
the product. Even in this case the �rm can survive if its initial capital is enough to
cover its costs in each period. Then, under a continual zero demand q(n; i; x) = 1 if
x + �Ci > Hn. However, at the other extreme if the demand goes up to M , and we
have no stock in inventory, then we will also have to pay the shortage cost rM . Hence
in general q(n; i; x) = 1 if x > (H + rM)n� �Ci.
Looking at the situation when we cannot survive, we should investigate the most
advantageous run of demands. To avoid shortage costs the demand should not be bigger
than the inventory in stock, and so the best we can hope for is to have a demand of
exactly the same amount of the items in stock. Each period the �rm only makes
a pro�t, and hence improves its �nancial position if the amount ordered is at least
k� = bH + w=(S � C)c + 1 (so that (S � C)k� > H + w, the pro�t is bigger than the
7
costs) where bxc denotes the integer part of x. Thus, if we start with zero components
in stock, we can only hope to survive if we can order k� or more components initially
(i.e. x > Ck�+H +w). Thus, if x < Ck
�+H +w there is no chance of survival. If we
start with 0 < i < k� components in stock, then the only di�erence is that the money
available to us at the start is x+�Ci, if we sell i items at the end of the period we have
a capital of x+Si�Ck��H�w and hence we will not survive if x < Ck
�+H+w�Si.
Let us now focus on the properties that relate the capital available with the stock
in inventory.
Lemma 2
i) For any i � M , q(n; i; x) � q(n; i+ j; x� Cj � w) 8j � 0, and x+ �Ci � 0.
ii) For i < M q(n; i + j; x) � q(n; i; x+ (S + r)j), and for i � M , q(n; i+ j; x) �
q(n; i; x+ Sj) 8j � 0, and x+ �Ci � 0.
Proof:
i) Suppose the optimal action in state (n; i + j; x � jC � w) is
k(n; i+ j; x� jC �w) = Æ � 0, and consider the particular action that orders j + Æ in
state (n; i; x). As i � M , from equation (1) we have
q(n; i; x) �MXd=0
p(d) q(n� 1; (j + Æ) + i� d; x+ Sd� C(j + Æ) +
�min(w; (j + Æ)w)�H); from Lemma 1 ii)
�MXd=0
p(d) q(n� 1; (i+ j) + Æ � d; x� Cj � w + Sd� CÆ
�min(w; Æw)�H)
= q(n; i+ j; x� Cj � w)
as �min(w; (j+Æ)w) � �w�min(w; Æw) 8j; Æ; w � 0, arriving at the desired inequality
q(n; i; x) � q(n; i+ j; x� jC � w).
ii) It is suÆcient to prove property (ii) for j = 1, and the proof will be by induction
on n. The result is trivially true for n = 0 as q(0; i; x) = 1 8i � 0 and x + �Ci � 0.
Assume that it is true for n� 1. Let I(i) =
8<:
0 if i � 0
1 if i > 0, and let k(n; i+ 1; x) = Æ
be the optimal action in state (n; i + 1; x), and consider the action that orders Æ in
8
state (n; i; x + S) for i > M , or state (n; i; x+ S + r) for i � M . We have
q(n; i+ 1; x) =
iXd=0
p(d)q(n� 1; i + 1 + Æ � d; x + Sd� CÆ �min(Æw; w)
�rmax(0; d� (i + 1))�H) +
MXd=i+1
p(d)q(n� 1; Æ; x+ S(i+ 1)
�CÆ � r(d� i� 1)�min(Æw; w)�H)
�iX
d=0
p(d)q(n� 1; i+ Æ � d; x+ S + Sd� CÆ �min(Æw; w)
+rI(M � i)�H) +
MXd=i+1
p(d)q(n� 1; Æ; x+ S + Si
�CÆ � r(d� i) + rI(M � i)�min(Æw; w)�H)
� q(n; i; x+ S) if i � M; and is
� q(n; i; x+ S + r) if i < M
where the second inequality follows from the induction hypothesis, and Lemma 1 ii).
This proves the induction hypothesis holds for n, and the result follows.
�
In other words, when i � M , we prefer to have Cj + w capital instead of j com-
ponents in stock, on the other hand for an inventory level i > M we prefer to have
Sj extra cash in capital than j units in stock, and for i � M we prefer Sj + r extra
capital than j items in stock.
We now consider the probability of the �rm surviving over a in�nite horizon. q(i; x) =
limn!1
q(n; i; x) is the probability that the �rm will survive forever given that it has i
components in stock and x units of capital. We describe the properties of the function
q(i; x) in the following Lemma and Theorem.
Lemma 3
i) q(i; x) = limn!1
q(n; i; x) exists.
ii) For any i � M , q(i; x) � q(j + i; x� jC) 8j � 0, x+ �Ci � 0.
iii) For i > M , q(i+ j; x) � q(i; x + jS), and for i � M
q(i+ j; x) � q(i; x+ jS + r) 8j � 0, x + �Ci � 0.
9
iv) q(i; x) is non-decreasing in x.
v) q(i; x) is non-decreasing in i.
Proof:
q(n; i; x) is bounded above by 1 and below by 0, and from Lemma 1 i) is monotonic
non-increasing in n. As bounded monotonic sequences converge, point i) follows. Points
ii) and iii) follow immediately by taking the limit in the results of Lemma 2, and points
iv) and v) follow by taking the limit in the results of Lemma 1 parts ii) iii).
�
Theorem 2
i) q(i; x) is non-decreasing in �.
ii) q(i; x) is non-decreasing in S.
iii) q(i; x) is non-increasing in C.
iv) q(i; x) is non-increasing in H.
v) q(i; x) is non-increasing in r.
vi) q(i; x) is non-increasing in w.
Proof:
From Lemma 3 we know that the q(i; x) exists. Hence, points i)-vi) follow immedi-
ately by taking the limit in the results of Theorem 1 i)-vi).
�
However we still have to prove that these results are non-trivial, namely that there
are levels of capital x where survival is a real possibility, i.e q(i; x) 6= 0.
Theorem 3 For all inventory levels i, q(i; x) > 0 for some �nite x.
Proof:
Consider � = 0 (that there is no loan), and a policy that orders up to 2M each
period. Such a policy does not incur any shortage cost, except perhaps in the �rst
period. Hence, Theorem 1 of Archibald et al1 shows that with an initial capital of
2MC +H +w+ rM there is a positive probability of survival if we take the overhead
costs for the reduced model to be H 0 = H+w (note that having an overhead cost of H
10
and an ordering cost of w ensures that the �rm has at least as much capital as when
the overhead cost is H 0). Hence, the result is true for � = 0.
On the other hand if � > 0, because of Theorem 2 i) the result also holds, as more
capital is available through the loan.
�
Experimental Results
We present a set of comparisons for the behaviour of the model under di�erent values
of �, i, w, and r. In the example the holding cost is H = 10, the cost of one component
is C = 3, and the selling price of each item produced is S = 5, and we have a Poisson
demand process with mean 7.5 truncated at 20 (i.e the probability that the demand is
higher than 20 is added to the probability of the demand equals 20).
First we analyse the e�ect di�erent colateral levels � have on the probability of
survival and the ordering policy. We consider a case where the ordering and shortage
costs are zero. In Figure 2 we present, for a given inventory level i = 3, a graph
comparing the probability of survival as capital increases, and a graph comparing the
ordering policy with di�erent levels of capital.
Note in the �rst graph of Figure 2 how the probability of survival increases with
bigger loans (better � valuation by the bank of the inventory in stock). This is what
we expected (recall Theorem 1 i)). With a capital of x = 13 and no loan (� = 0)
the survival probability is 0:34, whereas with � = 0:3 the probability of survival is
0:74, and with � = 0:8 it is 0:99. We need less capital to start surviving when the
colateral level � is higher. Hence, for a capital of x = 3 there is a probability of 0:95 of
surviving with � = 0:8, but there is no chance of surviving in the long run for � = 0:3.
This con�rms Buzacott and Zhang4 who found that the bank's assessment of the risk
associated with the �rms inventory had a large impact on its pro�tability. In our case
it has a large e�ect on the probability of survival in the long run.
Archibald et al1 found that as the capital level increased we stopped having a unique
policy - single k policy - and we started to have multiple ordering policies - several
values for k - that maximised the survival probability. A parsimonious policy is one
that orders the least amount possible - the smallest of such k values. We show in the
second graph of Figure 2 the parsimonious policy (and the point T where we start
11
Figure 2: Survival probability and Optimal Policy for di�erent � loans
T=point at where there starts to be more than one optimal policy
12
having multiple policies). Note in the second graph of Figure 2 how the �rm cannot be
too conservative in its ordering policy, and how we need for small quantities of captial
to order at least k� = 6 components to survive (i.e for q > 0), for any � level. This
graph also shows that the order policy is non-monotonic with respect to capital. For
example, for � = 0:3 and capital x = 13 the unique optimal ordering policy is k = 8;
however for x = 14, the unique optimal ordering policy is k = 7. This is because
of short term e�ects of what size of demand next period guarantees survival for the
next two periods at least. It also seems that for bigger loans the �rm is willing to buy
more components; the parsimonious policy for � = 0 buys at most 13 items, where
as for � = 0:8 the parsimonious policy buys at most 16 items. However, this is not
true for any given capital (and inventory) level; for x = 24 with � = 0 the optimal
policy is k = 8 whereas with � = 0:3 the optimal policy is k = 7. We cannot state
that the ordering policy will always be bigger for bigger � levels (when x, and i are
�xed). Again this is caused by a short term e�ect of trying to maximise the demand
next period which ensures survival over the next two periods. On the other hand, the
ordering policy is not more than k = 16 components, this is reasonable as we have
i = 3 components in stock, and the maximum demand possible in one period is 20,
and there is no shortage cost.
In Figure 3 we show the behaviour of the probability of survival and the ordering
costs for di�erent inventory levels i = 3; 5; and 7. In this case we are considering a
shortage cost of r = 0:5, and a loan of � = 0:5. In the �rst graph of Figure 3 we can
see how the probability of survival increases with larger inventory levels. This is also
what happened when the colateral for the loan was bigger. For higher inventory levels
of stock the ordering policy tends to be smaller. Note how for i = 3 we do not order
more than k = 22 components (ordering exactly that amount for x = 62; 64), where
as for i = 3 we do not order more than k = 13 components. However, we cannot say
that the ordering policy will always be smaller for bigger i inventory levels (when x,
and � are �xed). For example two cases where the policy for i = 3 is smaller than for
i = 5 and i = 7 are: 1) for capital x = 9, the optimal policy for i = 3 is k = 7, but
for i = 5 it is k = 9, and 2) for capital x = 4, the optimal policy for i = 3 is k = 6,
but for i = 7 it is k = 7. Note how the ordering policy for i = 3 can be bigger than
17 items, which did not happen before (second graph Figure 2) when there were no
shortage cost. In this case as there is a shortage cost we expect the �rm to order more
13
to cover a possible high demand in a given period.
To analyse the e�ect of di�erent ordering costs, we have tried w = 0; 3; 30 with two
di�erent capital levels x = 30; 50. In Figure 4, we present a graph showing the changes
in the survival probability and optimal policy as the inventory increases. In the �rst
graph we have not included the results for w = 0; x = 30, as it is very similar to
w = 3; x = 50, and the results for w = 0; x = 30, which gave a line that almost went in
q = 1 across for all i inventory levels. Note how the probability of survival increases as
the inventory level increases. It is clear that a smaller ordering cost produces a higher
probability of survival. In the second graph of Figure 4 we show the ordering policy
as the inventory increases. Note how for very high ordering costs w = 30 (three times
the overheads) the company will try not to order, and so it needs high initial inventory
levels to survive (e.g. i > 30). The company will also try to order more components
when w = 3 than when no ordering costs are involved (i.e w = 0).
To compare the e�ect di�erent shortage costs have on the survival probability and
the policy, we have tried r = 1 (13of the component cost), r = 3 cost of the one
component, r = 10 the cost of the overhead costs. We show the results in the graphs of
Figure 5 with changes in capital. Again, we can see how the probability increases with
smaller shortage costs. Note how for a shortage cost of r = 10 there is a probability of
surviving of q < 0:01, where as for the other costs it is q � 0:98, there is a considerable
di�erence. We have that the parsimonious policy orders no more than k = 16 for r = 1
for any i, and x capital level, where as for r = 3 it is no bigger than k = 19, but in both
cases it decreases when we have enough capital available (e.g. x > 106, and x > 145
respectively). On the other hand, the parsimonious policy for r = 10 continues to
increase up to k = 20 for these capital levels. This is reasonable, as the �rm will tend
to avoid any shortage cost, as one unsold item in this case will be as expensive as the
overhead costs of the period.
14
Figure 3: Survival probability and ordering policies for di�erent inventory levels.
T=point at where there starts to be more than one optimal policy
15
Figure 4: Survival probability and ordering policies for di�erent ordering costs.
16
Figure 5: Survival probability and ordering policies for di�erent shortage costs.
17
Average Reward Model
In this section we analyse the model for an established manufacturing �rm. We assume
that the objective for such a company is to maximise the average reward (pro�t) per
period. In this case there is no constraint on the amount of capital as it is assumed
that the �rm has enough to �nance any purchase it wants. Thus, the state of a �rm
at the start of any period is described completely by the number of components in
stock. This is a countable state, �nite action, unichained Markov decision process.
Hence, the standard results for the average reward Markov decision processes hold (see
Puterman6).
Let g be the average reward per period under the inventory policy that optimises
the average reward period and let v(i) be the bias term of starting with i items in
stock. The optimality equation of the dynamic programming model for the �rm under
the given assumptions is
g + v(i) = maxk
fMXd=0
p(d)( Smin(i; d)� kC � rmax(0; d� i) (2)
�min(w; kw)�H + v(i+ k �min(i; d) )g
Let k(i) = argmaxfP
M
d=0 p(d)(Smin(i; d) � kC � H � v(i + k � min(i; d))g be the
optimal policy for a given inventory level i.
We assume there is a limit N on the components the �rm may buy in any given
period, which represents the constraint on the available supply on the market. N can
be considered much larger than M (at least 2M < N). Recall that a parsimonious
policy is one that orders the least amount possible to maximise the survival probability.
Our aim is to compare the parsimonious policy for the average reward model with that
of the survival probability introduced in the second section.
Lemma 4 The optimal average reward and bias terms that satisfy the model stated in
equation (2) are
g = (S � C)d�H �wd
N;
v(i) =
8>>>><>>>>:
Ci+ (S � C)d+ iw
Ni � M;
Si� (S � C)iP
d=0
(i� d)p(d)� rP
M
d=i+1(d� i)p(d) i < M;
+P
i
d=0 p(d)(i� d)wN
(3)
18
where the parsimonious policy is given for w > 0 by
k(i) =
8<:
0 if i � 2M
N if i < 2M;
(4)
and for w = 0 by
k(i) =
8>>><>>>:
0 if i � 2M
2M � i if M < i < 2M
M if i � M:
(5)
Proof
The proof uses the policy iteration algorithm for dynamic programming models (see
Puterman6). Let us �rst focus on the parsimonious policy described in equation (4) for
w > 0. When this policy is applied, the average reward and bias terms, equation (2),
satisfy the following equation.
g + v(i) =
8>>>>>>>>><>>>>>>>>>:
MPd=0
p(d)(Sd�H + v(i� d)) i � 2M
MPd=0
p(d)(Sd�NC � w �H + v(i+N � d) M � i < 2M
MPd=0
p(d)(Smin(i; d)�NC � rmax(0; d� i) i < M
�w �H + v(i+N �min(i; d))
;
It is easy to verify by substitution that the values of g and v(i) in equations (3) satisfy
this equation.
Note that if we describe the two expression for v(i) in equation (3) as v1(i) if i � M ,
and v2(i) if i < M , then expression v1(i) � v2(i) since
(S � C)
iXd=0
(i� d)p(d) > (S � C)(i� d):
Now apply a policy improvement step to verify that the policy of equation (4) is
optimal.
For states i � 2M , and w > 0 the policy improvement step looks for the action k
which maximises
MXd=0
(Sd� kC �min(w; kw)�H + v(i+ k � d))p(d);
19
and we wish to show that this occurs when k = 0. This would be the case if
Sd�H + v(i� d) � Sd� kC � w �H + v(i+ k � d) for all k:
Since i� d and i + k � d � M the corresponding bias term v(i� d), and v(i+ k � d)
are in fact v1(i� d) and v1(i + k � d) so we require Sd�H + C(i� d) + (S � C)d +
(i � d)wN� Sd � H + C(i + k � d) + (S � C)d + (i + k � d)w
N. Simplifying we need
(i� d)wN� (i+ k � d)w
N� w. Since k < N this inequality holds.
For M � i � 2M , we want to show that
MXd=0
(Sd�NC�w�H�v(i+N�d))p(d) �MXd=0
(Sd�kC�min(w; kw)�H+v(i+k�d))p(d)
for any k, 0 � k � N . If i+ k� d � M the corresponding bias term (v function) is v1.
This means we want
Sd�NC � w �H + C(i+N � d) + (S � C)d+w
N(i+N � d) �
Sd� kC � w �H + C(i + k � d) + (S � C)d+w
N(i+ k � d):
This is trivial as k � N . On the other hand if i+ k� d < M , then the right hand side
is v2(i+ k � d) not v1(i+ k� d) and since v2(i+ k � d) � v1(i+ k � d) the inequality
still holds.
For i � M , we want to show that
MXd=0
p(d)(Smin(i; d)�NC � w �H � rmax(0; d� i) + v1(i+N � d)) �
MXd=0
p(d)(Smin(i; d)� kC � rmax(0; d� i)� w �H + v(i+ k �min(i; d)):
This reduces to the previous case since the shortage r, and the sales S terms cancel
on both sides and v(i+ k �min(i; d)) � v1(i + k �min(i; d)).
Now look at the parsimonious policy for w = 0, and i � 2M . First evaluate the
policy described by equation (5). When the policy is applied, the average reward and
bias terms satisfy the following equation.
g + v(i) =
8>>><>>>:
PM
d=0 p(d) (Sd� (2M � i)C �H + v(2M � d)) for i > M
PM
d=0p(d) (Smin(i; d)�MC � rmin(0; d� i)�H
+v(i+M �min(i; d))) for i � M
20
It is easy to verify by substitution that the values of g and v(i) in equations (3) satisfy
this equation.
Now apply a policy improvement step to verify that the policy is optimal. For i > M
the policy improvement step looks for the action k which maximises
MXd=0
(Sd� kC �H + v(i+ k � d))p(d):
Since
v(i+ 1)� v(i) = S
MXd=i+1
p(d) + C
iXd=0
p(d) + r
MXd=i+1
p(d) > C if i < M;
and v(i+1)�v(i) = C if i � M , this expression is maximised when k is chosen so that
i+ k � d � M for all possible values of demand, d. Hence any k � 2M � i is optimal.
For i � M the optimal policy improvement step looks for the action k that maximises
iXd=0
p(d)(Sd� kC �H + v(i+ k � d)) +
MXd=i+1
p(d)(Si� kC � r(d� i)�H + v(k))
Since v(i+1)� v(i) > C if i < M , and v(i+1)� v(i) = C if i � M , this expression is
maximised when k is chosen so that k � M . Hence, the policy given by equation (5)
is optimal.
�
Note that the policy when w > 0 suggest that the company should buy as many
components as are available in the market if the inventory level is smaller than 2M .
The �rm is trying to avoid the payment of ordering costs as much as possible. On the
other hand when there is no ordering cost w = 0 we will only order up toM components
if the inventory level is less than M , and no more than 2M � i if the inventory level is
bigger than M . As there is no ordering cost the company orders whenever it is needed,
depending on the inventory in stock, and not more than 2M components each period.
We now compare both the pro�t maximising model and the survival probability
one. We show that a company should be more cautious and risk averse in its `survival'
phase than in its mature `pro�t maximising' phase. Caution means ordering fewer
components. Buying a large number of components depletes the capital reserves, but
gives the chance of high pro�ts if the demand is high (as well as reduce the shortage
cost). On the other hand, buying fewer items depletes the reserves less, but one gives up
21
the chance of higher pro�ts in case the demand is high in the next period. The following
theorem shows that the parsimonious policy for the survival probability function is
more cautious than the parsimonious policy for the average reward if w > 0. Let
k0(n; i; x) and k
0(i; x) be the parsimonious policy for the survival probability model in
the �nite, and in�nite horizon respectively. That is the smallest optimal k 2 k(n; i; x),
and k 2 k(i; x) (i.e. k0(n; i; x) = minfk(n; i; x)g, k0(i; x) = minfk(i; x)g). Likewise let
k0(i) be the parsimonious policy for the average reward model.
Theorem 4 If w > 0 then
i) k0(n; i; x) � k
0(i) for all n; i, and x.
ii) k0(i; x) � k
0(i) for all i, and x.
Proof
i) Since for i � 2M , k0(i) is N the largest possible order, then trivially
k0(n; i; x) � k
0(i), and k0(i; x) � k
0(i).
De�ne
Q(k; n; i; x) = fMXd=0
p(d) q(n� 1; i+ k �min(i; d); x+ Smin(i; d)� Ck
�rmax(0; d� i)�min(w; kw)�H)g:
For i � 2M ,we need to show that in the survival probability case, the optimal order
is zero. Suppose we take an order of size Æ, then
Q(Æ; n; i; x) =
MXd=0
p(d) q(n� 1; i+ Æ � d; x+ Sd� CÆ � w �H)
�MXd=0
p(d)q(n� 1; i� d; x+ Sd� w �H)
�MXd=0
p(d)q(n� 1; i� d; x+ Sd�H) = Q(0; n; i; x)
where the �rst inequality holds from Lemma 2 i), and the second form Lemma 1 ii).
Hence k(n; i; x) = 0 � k(i).
ii) The proof that k(i; x) � k(i) follows in exactly the same way using Lemma 3 instead
of Lemma 2.
�
22
The results also hold in the case w = 0, from the proofs in Archibal et al1. Hence,
Theorem 5
If w=0, then
i) k0(n; i; x) � k
0(i) for all n; i, and x.
ii) k0(i; x) � k
0(i) for all i, and x.
Conclusion
We have presented three di�erent extensions to the model presented by Archibald et al1
and highlighted their validity in di�erent ways.
We have shown that to maximise the probability of survival start-up companies
should employ more conservative strategies for ordering components parts than more
established �rms. We have shown the in uence the di�erent costs have on the probabil-
ity of survival and in the ordering policy. These include how having as big as possible
colateral for the loans increases the probability of survival, and how the shortage and
ordering cost translates into bigger orders, and smaller probabilities of survival.
It will also be of interest to consider further extension to the model to include
marketing and technology decisions, as well as the case of several components and
products.
References
1 Archibald TW, Thomas LC, Betts JM and Johnston RB (1997). Should start-up
companies be cautious ? Inventory policies which maximise survival probabilities.
Preprint paper, University of Edinburgh.
2 Modigliani F and Miller MH (1958). The Cost of Capital, Corporation Finance
and the Theory of Investment. The American Economic Review 48: 261-297.
3 Cifarelli MD, Masciandaro D, Peccati L, Salsa S and Tagliani A (2002). Sucess or
failure of a �rm under di�erent �nancing policies: A dyanmic stochastic model.
Eur J Opl Res 136: 471-482.
4 Buzacott JA and Zhang RQ (2001). Inventory Management with Asset-Based
Financing. Working paper, Schulich Shool of Business, York University.
23
5 Freel MS (1999). The �nancing of small �rm product innovation within the UK.
Technovation 19: 707-719.
6 Puterman ML (1994). Markov Decision Processes: Discrete Stochastic Dynamic
Programming. John Wiley & Sons: New York.
24